TopolearnSavedModel is the format TF Serving, LiteRT conversion and most Google Cloud tools consume.
A SavedModel is a directory containing the computation graphs (as tf.function traces), the variable values and signatures: named entry points with fixed input specs, like serving_default. Loading it does not need the Python class that created it, which makes it the hand-off format for production.
Two ways to produce one:
- For any
tf.Module, calltf.saved_model.save(obj, path). Methods decorated with@tf.function(input_signature=...)are saved; if exactly one exists, it becomes the default signature. You can passsignatures=explicitly. - For Keras 3 models, use
model.export(path), which writes an inference-only SavedModel with aserveendpoint. (Usemodel.save("m.keras")for the Keras format if you want to keep training later.)
tf.saved_model.load(path) returns an object whose saved functions and variables are restored, and loaded.signatures["serving_default"] gives a function taking keyword tensors and returning a dict.
Inspect a SavedModel from the shell with saved_model_cli show --dir path --all, which lists signatures with input and output names, shapes and dtypes. Always check this before deploying, since clients must send exactly those inputs.
Example
import tensorflow as tf
class Scaler(tf.Module):
def __init__(self):
super().__init__()
self.w = tf.Variable(2.0)
@tf.function(input_signature=[tf.TensorSpec([None], tf.float32)])
def __call__(self, x):
return x * self.w
tf.saved_model.save(Scaler(), "scaler_model")
loaded = tf.saved_model.load("scaler_model") # no Scaler class needed
print(loaded(tf.constant([1.0, 3.0])).numpy()) # [2. 6.]
print(list(loaded.signatures.keys())) # ['serving_default']
# Keras 3 models: model.export("export_dir") -> SavedModel for serving
# Inspect from the shell: saved_model_cli show --dir scaler_model --allRemember
- SavedModel = graphs + variables + signatures, no Python class needed
tf.saved_model.savefor modules,model.exportfor Keras models- Check signatures with
saved_model_cli show
Watch out for
- Saving a tf.function without an input signature and finding no serving signature in the export.